The problem
An IT team’s tickets (incidents, changes, requests and tasks) come out of ServiceNow as a set of exports. Questions like “which teams have the most open incidents?” or “what emergency changes happened this week?” should have an answer anyone on the team can get in seconds, without building a spreadsheet first.
How it works
- You ask a question in plain English, in the chat or in the bar at the top.
- Claude turns it into a structured query. OneTouch checks what comes back, and Cube, a data layer over PostgreSQL, runs it.
- Every query only covers the teams that person is allowed to see.
- The answer comes back as a table you can download in eight formats, or as a chart. There are also ten ITIL dashboards in Metabase, and dashboards people build themselves.
- An Architecture page draws the whole system in 2D and 3D, with live health checks and a risk monitor that turns a block red when something is about to go wrong.
Decisions I made
- The AI never touches the database. Claude describes the query, OneTouch checks it, and Cube runs it.
- Plan for the internet failing in the middle of a live presentation. Answers are saved, and an offline switch lets the assistant answer from them while everything else keeps running on one computer.
- A daily AI budget and a usage report, so the cost is visible before it becomes a problem.
- Accounts with passwords, visibility by team, and a lockout after repeated wrong passwords.
The hard part: keeping the documentation true
The README explains every feature, how to run it and how to rebuild it from scratch, with every file printed in full. A check fails as soon as the README no longer matches the code, so it can’t go stale without anyone noticing.
Keeping it live: DevOps and reliability
OneTouch had to stay up for a live presentation to an enterprise client, served from home hardware with no cloud budget. I took it from “runs on my laptop” to a monitored, publicly hosted service with a rehearsed failover plan.
- Containerized five services (the Node.js app, PostgreSQL, Cube, Metabase and a Cloudflare Tunnel) with Docker Compose. The site is public on a custom domain with no inbound ports open, and every service listens only on localhost.
- Migrated production from a laptop to a dedicated host. I backed up the database and dashboard volumes and restored them before the first start, because a first start on empty volumes would have rebuilt the database and lost everyone’s assignments.
- Built active/standby failover with a split-brain guard. Both machines share one tunnel token, so if both served, Cloudflare would split visitors between two databases. A PowerShell watchdog reads Cloudflare’s status (530 means no tunnel is connected) and a host-identity header on every response, so only one machine ever serves.
- The standby takes over with one command, but only after the site has had no connector for a full minute: longer than the 30 seconds after which the primary drops its own tunnel when it loses internet, so a brief blip can’t lead to two tunnels.
- It recovers on its own. Auto sign-in, Docker restart policies and the watchdog bring the whole site back after a reboot or a power cut, and the watchdog restarts the tunnel if it dies.
- An external uptime check watches the public URL, because a machine without power can’t report itself. The watchdog also sends phone alerts when the tunnel won’t start, another host is serving, the internet drops or the app keeps failing.
- GitHub Actions runs build checks, unit tests, a dependency audit and a smoke test against the live site. I rotated secrets that had been committed to git history, gave Cube a read-only database role, and added security headers (CSP, HSTS).
Rehearsed end to end: I cut the primary’s power at the wall and the standby took over. When the primary came back, it kept its tunnel off because the standby was serving. Once the standby let go, the primary was serving again about 15 seconds after it noticed.
- 5services in one Docker Compose stack, with no inbound ports open
- 18 sfrom the tunnel dying to serving again, with no one touching it
- ~1.5 minto come back after a reboot or power cut, with no one at the machine
- ~1 minof downtime to move hosts, with every table’s row count matching
How I worked: I set the goals and constraints, made the decisions, and directed AI coding tools through the implementation. I ran and verified every rehearsal myself.
In numbers
- 8download formats for every answer
- 10ITIL dashboards
- 2D + 3Dmaps of its own architecture
- 6risks watched live, from disk space to password guessing
OneTouch was built for a real IT team, so the portfolio demo uses generated data and prepared answers in place of Claude. No real tickets, names or client details appear in it.